Glucometer accuracy and implications
for clinical studies
AJ Le Compte, CG Pretty, GM Shaw, JG Chase
Example research study
Introduction
3.5
Elucidating links between glycemic control and clinical outcome
requires reliable discrimination between groups with different
target blood glucose (BG) cut-offs. Point-of-care glucometers are
commonly used, but lower accuracy means BG errors will
impact classification and thus outcome analyses. This study
reanalyses a BG control trial with an error model of a typical
glucometer to assess the impact of sensor errors on
interpretation of trial results.
Research question:
Classify patients into “good” and
“poor” BG control categories
Distribution of cTIB in clinical
results
3
• “Good” = at least 70% of BG
measurements in target BG band
(4 – 7 mmol/L, 72 – 126 mg/dL)
2.5
“Good”
control
“Poor”
control
Density
2
• “Poor” = less than 70% of
measurements in target band.
1.5
Clinical results
• 76% were “Good”
• 24% were “Poor”
1
Methods
Glucometer used for clinical
BG measurements.
0.5
0
0
0.1
0.2
10
BG (mmol/L)
0.4
0.5
0.6
Cumulative time in band
0.7
0.8
Start: Clinically reported
time in target band: 85%
12
BG profiles from N=301 patients (stay>24hrs) from the SPRINT
trial with BG measurements measured using the Arkray
SuperGlucocard II GT-1630 were reanalysed to determine how
sensor errors affect classifying patients into categories of “good”
and “poor” control.
0.3
0.9
What is the chance a “good”
patient was actually “poor”?
1
Model of BG sensor to
estimate true glucose profile
8
Clinical accuracy
4
The defined cut-off for “Good” control for a patient was at least
70% of BG in 72–126mg/dL (cTIB>=0.7), and “Poor” as less than
70% (cTIB<0.7), based on original observed clinical BG. The
number of true BG profiles that resulted in misclassification
between “Good” and “Poor” control for a patient was recorded.
6
12
18
24
30
36
42
Sensor model
48
4
Result: Patient very likely
to be classified “good”
despite glucometer errors
12
BG (mmol/L)
10
2
0
Chance of mis-classification
A “deadband” around the
cut-off would reduce this
misclassification.
6
12
18
24
30
36
42
Deadband of uncertain patient
results due to glucometer errors
3.5
3
Chance of misclassification < 10%
42
48
Time (hours)
Repeat 100x to
generate range of
potential “true” BG
48
For cTIB cut-off of 0.5 few SPRINT patients are below cutoff, and those that are were close to 0.5. Thus this cut-off
value may result in less reliable interpretation of results.
3
Distribution of cTIB in clinical
results
2.5
91%
88%
2.5
Density
Patients unlikely to be influenced by
glucometer errors
2
1.5
1
1.5
0.5
0
1
Chance of misclassification < 10%
99%
67%
0
0
0.1
0.2
0.3
0.4
0.5
0.6
Cumulative time in band
0.7
0.8
0.9
0
0.1
0.2
0.3
0.4
0.5
0.6
Cumulative time in band
0.7
0.8
0.9
Glucometer errors could cause these patients to be
1placed in the wrong group. Thus there is greater
uncertainty in which group these patients belong.
Conclusions
1) Glucometers can distinguish between
patients that received good and poor BG
control.
2) Cut-off values must account for sensor
errors. Deadbands may be required to
ensure patients are unlikely to be
misclassified.
0.6
0.4
0.2
0.2
0.4
0.6
0.8
Clinical reported time in band (cTIB)
1
1
Patients at risk of being placed in wrong group form a large
proportion of total cTIB < 0.5 patients.
0.5
0.8
0
0
36
Deadbands around cut-off values
1
Patients with cTIB near 0.7
were more likely to be
misclassified when
accounting for glucometer
error.
30
Time (hours)
If the “Good” cut-off was cTIB>=0.5 (95% of clinical patients)
then correct classification was 97% for good control patients, but
fell to 40% of poor control patients as most SPRINT patients with
cTIB<0.5 were close to the cut-off value and thus likely to be
affected by sensor error.
Good
Poor
24
4
83%
64%
Never
misclassified
97%
40%
18
6
2
Clinical
result
95%
5%
12
8
Density
76%
24%
6
Range of true time in
target band: 83-86%  all
above 70% cut-off
3.5
Good
Poor
6
2
0
Good control was clinically measured in 76% of patients (24%
with cTIB<0.7). Of these, 83% of “Good” and 64% of “Poor”
control would never be misclassified over all 100 runs due to
sensor error. The chance of misclassification was less than 10%
for 91% (good) and 87.5% (poor) of patients.
Never
misclassified
8
Time (hours)
Results
Clinical
result
Estimated true time in
target band: 83%
10
6
2
0
12
BG (mmol/L)
A model of sensor bias and variance (CV 2.7-3.5%, Regression:
y=3.92+0.97x) was used to estimate possible ‘true’ BG profiles
from measured/observed BG and repeated 100 times for each
patient.
3) The proportion of patients that could be
misclassified depends on the underlying
distribution. Hence, feasible cut-off values
are constrained to prevent sensor errors
resulting in significant misclassification.
4) Sensor errors affect the ability to observe
the population, and are independent of
Type-I and Type-II sampling errors.